submul_1.asm

a very popular packet of cryptography tools,it encloses the most common used algorithm and protocols

	fldd		24(up), %fr7
	xmpyu		%fr8R, %fr4L, %fr22
	xmpyu		%fr8L, %fr4R, %fr23
	xmpyu		%fr8R, %fr5L, %fr24
	xmpyu		%fr8L, %fr5R, %fr25
	xmpyu		%fr8R, %fr6L, %fr26
	xmpyu		%fr8L, %fr6R, %fr27
	fstd		%fr22, -0x78(%r30)	C mid product to  -0x78..-0x71
	xmpyu		%fr8R, %fr7L, %fr28
	xmpyu		%fr8L, %fr7R, %fr29
	fstd		%fr23, -0x70(%r30)	C mid product to  -0x70..-0x69
	xmpyu		%fr8R, %fr4R, %fr30
	xmpyu		%fr8L, %fr4L, %fr31
	fstd		%fr24, -0x38(%r30)	C mid product to  -0x38..-0x31
	xmpyu		%fr8R, %fr5R, %fr22
	xmpyu		%fr8L, %fr5L, %fr23
	fstd		%fr25, -0x30(%r30)	C mid product to  -0x30..-0x29
	xmpyu		%fr8R, %fr6R, %fr24
	xmpyu		%fr8L, %fr6L, %fr25
	fstd		%fr26, -0x58(%r30)	C mid product to  -0x58..-0x51
	xmpyu		%fr8R, %fr7R, %fr26
	fstd		%fr27, -0x50(%r30)	C mid product to  -0x50..-0x49
	addib,<>	-1, n, L(8_or_more)
	xmpyu		%fr8L, %fr7L, %fr27
	fstd		%fr28, -0x18(%r30)	C mid product to  -0x18..-0x11
	fstd		%fr29, -0x10(%r30)	C mid product to  -0x10..-0x09
	fstd		%fr30, -0x80(%r30)	C low product to  -0x80..-0x79
	fstd		%fr31, -0x68(%r30)	C high product to -0x68..-0x61
	fstd		%fr22, -0x40(%r30)	C low product to  -0x40..-0x39
	fstd		%fr23, -0x28(%r30)	C high product to -0x28..-0x21
	fstd		%fr24, -0x60(%r30)	C low product to  -0x60..-0x59
	fstd		%fr25, -0x48(%r30)	C high product to -0x48..-0x41
	fstd		%fr26, -0x20(%r30)	C low product to  -0x20..-0x19
	fstd		%fr27, -0x88(%r30)	C high product to -0x88..-0x81
	ldd		-0x78(%r30), p032a1
	ldd		-0x70(%r30), p032a2
	ldd		-0x38(%r30), p096b1
	ldd		-0x30(%r30), p096b2
	ldd		-0x58(%r30), p160c1
	ldd		-0x50(%r30), p160c2
	ldd		-0x18(%r30), p224d1
	ldd		-0x10(%r30), p224d2
	b		L(end1)
	nop

L(8_or_more)
	fstd		%fr28, -0x18(%r30)	C mid product to  -0x18..-0x11
	fstd		%fr29, -0x10(%r30)	C mid product to  -0x10..-0x09
	ldo		32(up), up
	fstd		%fr30, -0x80(%r30)	C low product to  -0x80..-0x79
	fstd		%fr31, -0x68(%r30)	C high product to -0x68..-0x61
	fstd		%fr22, -0x40(%r30)	C low product to  -0x40..-0x39
	fstd		%fr23, -0x28(%r30)	C high product to -0x28..-0x21
	fstd		%fr24, -0x60(%r30)	C low product to  -0x60..-0x59
	fstd		%fr25, -0x48(%r30)	C high product to -0x48..-0x41
	fstd		%fr26, -0x20(%r30)	C low product to  -0x20..-0x19
	fstd		%fr27, -0x88(%r30)	C high product to -0x88..-0x81
	fldd		0(up), %fr4
	fldd		8(up), %fr5
	fldd		16(up), %fr6
	fldd		24(up), %fr7
	xmpyu		%fr8R, %fr4L, %fr22
	ldd		-0x78(%r30), p032a1
	xmpyu		%fr8L, %fr4R, %fr23
	xmpyu		%fr8R, %fr5L, %fr24
	ldd		-0x70(%r30), p032a2
	xmpyu		%fr8L, %fr5R, %fr25
	xmpyu		%fr8R, %fr6L, %fr26
	ldd		-0x38(%r30), p096b1
	xmpyu		%fr8L, %fr6R, %fr27
	fstd		%fr22, -0x78(%r30)	C mid product to  -0x78..-0x71
	xmpyu		%fr8R, %fr7L, %fr28
	ldd		-0x30(%r30), p096b2
	xmpyu		%fr8L, %fr7R, %fr29
	fstd		%fr23, -0x70(%r30)	C mid product to  -0x70..-0x69
	xmpyu		%fr8R, %fr4R, %fr30
	ldd		-0x58(%r30), p160c1
	xmpyu		%fr8L, %fr4L, %fr31
	fstd		%fr24, -0x38(%r30)	C mid product to  -0x38..-0x31
	xmpyu		%fr8R, %fr5R, %fr22
	ldd		-0x50(%r30), p160c2
	xmpyu		%fr8L, %fr5L, %fr23
	fstd		%fr25, -0x30(%r30)	C mid product to  -0x30..-0x29
	xmpyu		%fr8R, %fr6R, %fr24
	ldd		-0x18(%r30), p224d1
	xmpyu		%fr8L, %fr6L, %fr25
	fstd		%fr26, -0x58(%r30)	C mid product to  -0x58..-0x51
	xmpyu		%fr8R, %fr7R, %fr26
	ldd		-0x10(%r30), p224d2
	fstd		%fr27, -0x50(%r30)	C mid product to  -0x50..-0x49
	addib,=		-1, n, L(end2)
	xmpyu		%fr8L, %fr7L, %fr27
L(oop)
	add		p032a1, p032a2, m032
	ldd		-0x80(%r30), p000a
	add,dc		p096b1, p096b2, m096
	fstd		%fr28, -0x18(%r30)	C mid product to  -0x18..-0x11

	add,dc		p160c1, p160c2, m160
	ldd		-0x68(%r30), p064a
	add,dc		p224d1, p224d2, m224
	fstd		%fr29, -0x10(%r30)	C mid product to  -0x10..-0x09

	add,dc		%r0, %r0, m288
	ldd		-0x40(%r30), p064b
	ldo		32(up), up
	fstd		%fr30, -0x80(%r30)	C low product to  -0x80..-0x79

	depd,z		m032, 31, 32, ma000
	ldd		-0x28(%r30), p128b
	extrd,u		m032, 31, 32, ma064
	fstd		%fr31, -0x68(%r30)	C high product to -0x68..-0x61

	depd		m096, 31, 32, ma064
	ldd		-0x60(%r30), p128c
	extrd,u		m096, 31, 32, ma128
	fstd		%fr22, -0x40(%r30)	C low product to  -0x40..-0x39

	depd		m160, 31, 32, ma128
	ldd		-0x48(%r30), p192c
	extrd,u		m160, 31, 32, ma192
	fstd		%fr23, -0x28(%r30)	C high product to -0x28..-0x21

	depd		m224, 31, 32, ma192
	ldd		-0x20(%r30), p192d
	extrd,u		m224, 31, 32, ma256
	fstd		%fr24, -0x60(%r30)	C low product to  -0x60..-0x59

	depd		m288, 31, 32, ma256
	ldd		-0x88(%r30), p256d
	add		climb, p000a, s000
	fstd		%fr25, -0x48(%r30)	C high product to -0x48..-0x41

	add,dc		p064a, p064b, s064
	ldd		0(rp), r000
	add,dc		p128b, p128c, s128
	fstd		%fr26, -0x20(%r30)	C low product to  -0x20..-0x19

	add,dc		p192c, p192d, s192
	ldd		8(rp), r064
	add,dc		p256d, %r0, climb
	fstd		%fr27, -0x88(%r30)	C high product to -0x88..-0x81

	ldd		16(rp), r128
	add		ma000, s000, s000	C accum mid 0
	ldd		24(rp), r192
	add,dc		ma064, s064, s064	C accum mid 1

	add,dc		ma128, s128, s128	C accum mid 2
	fldd		0(up), %fr4
	add,dc		ma192, s192, s192	C accum mid 3
	fldd		8(up), %fr5

	add,dc		ma256, climb, climb
	fldd		16(up), %fr6
	sub		r000, s000, s000	C accum rlimb 0
	fldd		24(up), %fr7

	sub,db		r064, s064, s064	C accum rlimb 1
	sub,db		r128, s128, s128	C accum rlimb 2
	std		s000, 0(rp)

	sub,db		r192, s192, s192	C accum rlimb 3
	sub,db		%r0, climb, climb
	sub		%r0, climb, climb
	std		s064, 8(rp)

	xmpyu		%fr8R, %fr4L, %fr22
	ldd		-0x78(%r30), p032a1
	xmpyu		%fr8L, %fr4R, %fr23
	std		s128, 16(rp)

	xmpyu		%fr8R, %fr5L, %fr24
	ldd		-0x70(%r30), p032a2
	xmpyu		%fr8L, %fr5R, %fr25
	std		s192, 24(rp)

	xmpyu		%fr8R, %fr6L, %fr26
	ldd		-0x38(%r30), p096b1
	xmpyu		%fr8L, %fr6R, %fr27
	fstd		%fr22, -0x78(%r30)	C mid product to  -0x78..-0x71

	xmpyu		%fr8R, %fr7L, %fr28
	ldd		-0x30(%r30), p096b2
	xmpyu		%fr8L, %fr7R, %fr29
	fstd		%fr23, -0x70(%r30)	C mid product to  -0x70..-0x69

	xmpyu		%fr8R, %fr4R, %fr30
	ldd		-0x58(%r30), p160c1
	xmpyu		%fr8L, %fr4L, %fr31
	fstd		%fr24, -0x38(%r30)	C mid product to  -0x38..-0x31

	xmpyu		%fr8R, %fr5R, %fr22
	ldd		-0x50(%r30), p160c2
	xmpyu		%fr8L, %fr5L, %fr23
	fstd		%fr25, -0x30(%r30)	C mid product to  -0x30..-0x29

	xmpyu		%fr8R, %fr6R, %fr24
	ldd		-0x18(%r30), p224d1
	xmpyu		%fr8L, %fr6L, %fr25
	fstd		%fr26, -0x58(%r30)	C mid product to  -0x58..-0x51

	xmpyu		%fr8R, %fr7R, %fr26
	ldd		-0x10(%r30), p224d2
	fstd		%fr27, -0x50(%r30)	C mid product to  -0x50..-0x49
	xmpyu		%fr8L, %fr7L, %fr27

	addib,<>	-1, n, L(oop)
	ldo		32(rp), rp

L(end2)
	add		p032a1, p032a2, m032
	ldd		-0x80(%r30), p000a
	add,dc		p096b1, p096b2, m096
	fstd		%fr28, -0x18(%r30)	C mid product to  -0x18..-0x11
	add,dc		p160c1, p160c2, m160
	ldd		-0x68(%r30), p064a
	add,dc		p224d1, p224d2, m224
	fstd		%fr29, -0x10(%r30)	C mid product to  -0x10..-0x09
	add,dc		%r0, %r0, m288
	ldd		-0x40(%r30), p064b
	fstd		%fr30, -0x80(%r30)	C low product to  -0x80..-0x79
	depd,z		m032, 31, 32, ma000
	ldd		-0x28(%r30), p128b
	extrd,u		m032, 31, 32, ma064
	fstd		%fr31, -0x68(%r30)	C high product to -0x68..-0x61
	depd		m096, 31, 32, ma064
	ldd		-0x60(%r30), p128c
	extrd,u		m096, 31, 32, ma128
	fstd		%fr22, -0x40(%r30)	C low product to  -0x40..-0x39
	depd		m160, 31, 32, ma128
	ldd		-0x48(%r30), p192c
	extrd,u		m160, 31, 32, ma192
	fstd		%fr23, -0x28(%r30)	C high product to -0x28..-0x21
	depd		m224, 31, 32, ma192
	ldd		-0x20(%r30), p192d
	extrd,u		m224, 31, 32, ma256
	fstd		%fr24, -0x60(%r30)	C low product to  -0x60..-0x59
	depd		m288, 31, 32, ma256
	ldd		-0x88(%r30), p256d
	add		climb, p000a, s000
	fstd		%fr25, -0x48(%r30)	C high product to -0x48..-0x41
	add,dc		p064a, p064b, s064
	ldd		0(rp), r000
	add,dc		p128b, p128c, s128
	fstd		%fr26, -0x20(%r30)	C low product to  -0x20..-0x19
	add,dc		p192c, p192d, s192
	ldd		8(rp), r064
	add,dc		p256d, %r0, climb
	fstd		%fr27, -0x88(%r30)	C high product to -0x88..-0x81
	ldd		16(rp), r128
	add		ma000, s000, s000	C accum mid 0
	ldd		24(rp), r192
	add,dc		ma064, s064, s064	C accum mid 1
	add,dc		ma128, s128, s128	C accum mid 2
	add,dc		ma192, s192, s192	C accum mid 3
	add,dc		ma256, climb, climb
	sub		r000, s000, s000	C accum rlimb 0
	sub,db		r064, s064, s064	C accum rlimb 1
	sub,db		r128, s128, s128	C accum rlimb 2
	std		s000, 0(rp)
	sub,db		r192, s192, s192	C accum rlimb 3
	sub,db		%r0, climb, climb
	sub		%r0, climb, climb
	std		s064, 8(rp)
	ldd		-0x78(%r30), p032a1
	std		s128, 16(rp)
	ldd		-0x70(%r30), p032a2
	std		s192, 24(rp)
	ldd		-0x38(%r30), p096b1
	ldd		-0x30(%r30), p096b2
	ldd		-0x58(%r30), p160c1
	ldd		-0x50(%r30), p160c2
	ldd		-0x18(%r30), p224d1
	ldd		-0x10(%r30), p224d2
	ldo		32(rp), rp

L(end1)
	add		p032a1, p032a2, m032
	ldd		-0x80(%r30), p000a
	add,dc		p096b1, p096b2, m096
	add,dc		p160c1, p160c2, m160
	ldd		-0x68(%r30), p064a
	add,dc		p224d1, p224d2, m224
	add,dc		%r0, %r0, m288
	ldd		-0x40(%r30), p064b
	depd,z		m032, 31, 32, ma000
	ldd		-0x28(%r30), p128b
	extrd,u		m032, 31, 32, ma064
	depd		m096, 31, 32, ma064
	ldd		-0x60(%r30), p128c
	extrd,u		m096, 31, 32, ma128
	depd		m160, 31, 32, ma128
	ldd		-0x48(%r30), p192c
	extrd,u		m160, 31, 32, ma192
	depd		m224, 31, 32, ma192
	ldd		-0x20(%r30), p192d
	extrd,u		m224, 31, 32, ma256
	depd		m288, 31, 32, ma256
	ldd		-0x88(%r30), p256d
	add		climb, p000a, s000
	add,dc		p064a, p064b, s064
	ldd		0(rp), r000
	add,dc		p128b, p128c, s128
	add,dc		p192c, p192d, s192
	ldd		8(rp), r064
	add,dc		p256d, %r0, climb
	ldd		16(rp), r128
	add		ma000, s000, s000	C accum mid 0
	ldd		24(rp), r192
	add,dc		ma064, s064, s064	C accum mid 1
	add,dc		ma128, s128, s128	C accum mid 2
	add,dc		ma192, s192, s192	C accum mid 3
	add,dc		ma256, climb, climb
	sub		r000, s000, s000	C accum rlimb 0
	sub,db		r064, s064, s064	C accum rlimb 1
	sub,db		r128, s128, s128	C accum rlimb 2
	std		s000, 0(rp)
	sub,db		r192, s192, s192	C accum rlimb 3
	sub,db		%r0, climb, climb
	sub		%r0, climb, climb
	std		s064, 8(rp)
	std		s128, 16(rp)
	std		s192, 24(rp)

	ldd		-0xb0(%r30), %r13
	ldd		-0xb8(%r30), %r12
	ldd		-0xc0(%r30), %r11
	ldd		-0xc8(%r30), %r10
	ldd		-0xd0(%r30), %r9
	ldd		-0xd8(%r30), %r8
	ldd		-0xe0(%r30), %r7
	ldd		-0xe8(%r30), %r6
L(done)
ifdef(`HAVE_ABI_2_0w',
`	copy		climb, %r28
',`	extrd,u		climb, 63, 32, %r29
	extrd,u		climb, 31, 32, %r28
')
	ldd		-0xf0(%r30), %r5
	ldd		-0xf8(%r30), %r4
	bve		(%r2)
	ldd,mb		-0x100(%r30), %r3
EPILOGUE(mpn_submul_1)